Hebbian Model Learning - Computational Explorations in Cognitive Neuroscience - page 116

Some studies assessing the impact of NMDA re-
ceptor blockade show preserved behaviorally mea-
sured learning despite NMDA blockade. Such stud-
ies are generally difficult to interpret, however, be-
cause NMDA-blocking drugs have significant other
effects on behavior (Keith & Rudy, 1990). In addi-
tion, any preserved learning observed in these studies
can be accounted for by many of the above complica-
tions to the simple NMDA story (e.g., by the role of
non-NMDA factors like voltage-gated calcium chan-
nels or mGlu receptors).

The Learner

The World

Model

Parameters

Visible
State

World State Projection Inverse Model State

Figure 4.3: Sketch of the objective of model learning — to
produce an internal model that represents the “important” fea-
tures of the world: structural regularities, natural laws, con-
stants, general characteristics (not the literal copy of the world
that might otherwise be suggested by the figure). This is com-
plicated because the true state of the world is “hidden” from
us — we only see a sequence of piecemeal collapsed projec-
tions of the world state, which we have to somehow invert to
produce the corresponding state in our model.

Thus, the story, at least at the level of biological mecha-
nisms, may end up being somewhat more involved than
the elegant NMDA-mediated associativity described
above.
Nevertheless, despite a number of unresolved com-
plications, the biological data suggests that associative
learning is probably occurring in the cortex. The gener-
ally uncertain state of this data only serves to highlight
the need for computational models to explore what kind
of synaptic modification rules lead to effective overall
learning, because these details appear to be difficult to
extract from the biology.

the impoverished nature of our sensory access to the un-
derlying structure of the world, and the other has to do
with the overwhelming amount of information delivered
by our senses. These may seem at first glance to be con-
tradictory, but they are not — our senses deliver a large
quantity of low-quality information that must be highly
processed to produce the apparently transparent access
to the world that we experience.
We will see that the quality problem can be remedied
by introducing appropriate a priori (i.e., built in from the
start) biases that supplement and organize the incom-
ing information. The information overload problem can
be addressed by biasing learning in favor of simpler or
parsimonious models that end up ignoring some kinds
of information in favor of representing enough relevant
information in a manageable form. As we will elabo-
rate, both of these techniques can be recognized in the
methods of scientific research, which can be viewed as
an explicit and deliberate extension of the model learn-
ing process, and has many of the same problems.
The problem with our access to the world via our
senses is that we only receive a series of relatively lim-
ited two-dimensional snapshots (and sound bites, etc.),
which can be thought of as projections of the very high-

4.3 Computational Objectives of Learning

In this section, we develop a computational-level moti-
vation for one general goal of learning and see how this
goal can be accomplished using some of the biologi-
cal mechanisms discussed in the previous section. We
call this goal model learning to emphasize the idea that
learning is directed toward developing internal models
of the world. Figure 4.3 shows an illustration of how
an internal model captures some of the important fea-
tures of the world, and also why this is a difficult thing
to do. The basic idea is that there is some kind of under-
lying structure or regularities (e.g., natural laws, con-
stants, general characteristics) of the world, and that this
should somehow be represented to function properly.
From our subjective experience of having a relatively
“transparent” knowledge of the physical structure of the
world, one might think that model learning is easy and
automatic, but there are two fundamental and difficult
problems faced by model learning. One has to do with